Algorithmic Bias

Speaker A: 00:00:00

Hello and welcome to the Cognitive Engineering Podcast produced by me, Fraser McGruer, for Aleph Insights. In this series of podcasts, we take a look at interesting topics and discuss what we think they tell us about analysis and decision-making. I'm here with Nick Hare and Peter Coghill of Aleph Insights, and also this week with our special guest, Jess Whittlestone. Nick, kick us off. Yeah, so I've known Jess for a few years now, and recently when we met up, she was telling me about some work she's been doing on the ethics, on the issue of sort of ethics and artificial intelligence. And one of the things that we're talking about is this issue of algorithmic bias, you know, about sort of artificial intelligences, forming beliefs or making decisions in ways that we perhaps find morally questionable. So I thought it'd be great to have her on to talk a bit more about that. Brilliant. Jess, can you just tell us a bit more about yourself? Just introduce yourself.

Speaker B: 00:00:53

Yeah, sure. So I've had quite a mixed background. So I'm an academic, but I do a lot of sort of engagement with policy issues too. And my PhD was in behavioural science, where I was really interested in kind of like rationality and decision-making and the ways that people are irrational. And through that, I did some work with the UK government on sort of applying insights from behavioural science to policymaking. But I now have switched and I'm working much more on sort of ethical and policy issues surrounding artificial intelligence, because I think it's become really clear over the last few years that these advances in AI and machine learning are going to be really important for society. And I want to try and sort of help figure out how to make sure that that's safe and beneficial. And I should sort of flag that algorithmic bias is not my area of sort of expertise, but it's something that I have been thinking about a bit. And me and Nick had some interesting conversations on.

Speaker A: 00:01:40

Okay, so sorry, probably quite a basic question. But first off, I mean, what are we talking about when we're talking about algorithmic bias? Can you give me some examples?

Speaker B: 00:01:48

Yeah, so there have been a couple of really high profile examples recently, which I think was what prompted this conversation. One of the highest profile one was the use of machine learning in predictive policing and predictive criminal justice. So machine learning algorithms were being used to predict which criminals were most likely to re-offend. And what came out of this, and there was this huge sort of high profile study, an article that basically found that these algorithms were more likely, they were biased towards black people in some specific ways. Or against. Against. And they were more likely to, I think black people were more likely to be wrongly categorized as likely to re-offend when they weren't likely to re-offend than white people. And what we were talking about is the fact that what these machine algorithms are doing is really helping us to better understand patterns in existing data. And a lot of that existing data kind of represents historical decisions. And so the conversation we were having was then asking, well, maybe some of the time we just don't actually want to know, like it's not useful for us to know, like better understand patterns in existing data in order to make better decisions. So in criminal justice, actually, in order to make better decisions about who we incarcerate or who we give parole to, is it helpful to know how those decisions have been made historically, if we know that those decisions historically have been subject to a lot of bias?

Speaker A: 00:03:20

Yeah. Is it possible that the results of what an algorithm might produce is something that we don't like? You know, that's, that's the question. Well, no, hold on. Is it something we don't like or is it something that is wrong? Yeah, right. Well, I mean, I think there's a, sorry, Jess, go on. Jess.

Speaker B: 00:03:37

So I think one thing that's important here is sort of what is the purpose of better understanding historical data or historical decisions? It might be really useful for us to know that historical decisions have been biased if the purpose is to sort of think about how to make them less biased in future. But if the purpose is, if we're just thinking about trying to better identify who is likely to re-offend, for example, understanding how those decisions have been made in the past is perhaps not very useful if there's bias. So, yeah.

Speaker A: 00:04:08

So I've got a question. Do they think that this, so this, this, is it fair to say that the algorithm was making incorrect judgments and that the incorrect, incorrectness was, you know, more, had a kind of pushed, pushed the decision makers towards refusing parole for black people who otherwise had exactly the same characteristics and were like in however you can measure it as likely to re-offend, but they were getting parole refused, whereas white people weren't. And if so, why, why was that happening?

Speaker B: 00:04:49

So yeah, I mean, I think this gets at how complicated some of these issues are. So I think the issue here in the parole case and in the re-offending cases, so a machine learning algorithm was given a ton of historical data on, on re-offending rates. And I think it is, is a fact that in that data, black people were perhaps in some statistical sense, more likely to re-offend than white people. But that doesn't mean that the issue is that the algorithm then picks up that feature and says, ah, skin color is a predictor of how likely someone is to re-offend. And that is a very different, a different thing is like, so the worry is that actually the reason that we see these patterns in the data is not because skin color is a predictor of likelihood of committing crime. It's because of historical biases in the past. It's the fact that perhaps police have been harsher on black people and so are more likely to pick them up. So yeah, so things that aren't in the data, it's proxying

Speaker A: 00:05:52

for things that aren't in the data that probably should be. If we could measure things like harshness of police treatment or whatever, if those things, if we could measure those,

Speaker B: 00:06:00

we could factor that out. Yeah. But I suppose the bigger issue here that I was thinking about is, if we, if we understand what machine learning is doing is sort of picking out patterns in historical data, when we think about sort of fair and ethical use of machine learning, maybe we need to step back a bit further and say, when is doing that actually useful? For what kinds of data sets? For what kinds of questions? For what kinds of problems? Rather than what I worry a little bit is happening at the moment is like, let's just throw machine learning at all of our problems because it's cool and it's fun. Yeah. So yeah,

Speaker C: 00:06:35

so we're trying to ask the question, is the future behavior of the system likely to be similar to past behavior of the system? And do we want it to be? Yeah. So that's, I mean, I sort of happened upon a sort of taxonomy for different types of bias in parts of analysis. And I don't know if this chimes with any literature. I haven't done much research in it, but this seems like this is all about bias in the data. But it seems like there's other types of bias. I mean, it's related, it's all related, overlapping, but there's also sort of bias in the analysis. And this includes what assumptions we're making. One assumption might be that actually, yes, this data we have models the world well, and we can therefore use it to predict how the world will behave in the future. It might also include like the choice of which data fields are used to build your model and what type of model you use. And there's also a bias in ultimately how that analysis is used to inform a decision. So that, for example, might include, well, do we A, listen to the analysis at all? Or do we just make it up? Do we just wet finger it? Or do we, how intelligently do we do it? Do we take into account all the caveats and assumptions that the analyst has made? Or do we ignore that stuff and just take the simple message and just take the oversimplified view of things? Is that a useful distinction to sort of break it apart like that?

Speaker A: 00:08:09

I think this seems, well, I mean, I guess, yeah, like is the data representing a biased system? Wrongly or incorrectly biased system? I mean, if it is, you can throw that out, right? Let's just say that's, let's assume that we're not talking about that kind of thing. Because let's say we're trying to predict what decisions judges will make. And judges are all a load of racists. Well, we'll just be predicting, you know, racist decisions. So we obviously don't want to be doing that, right? The problem comes when you are, when you there might be bias in the real world. So you know, that actually certain features do accurately predict, you know, certain kinds of negative health outcomes, say, you know, if you're looking at allocating health resources, you know, make that, you know, there are, there are situations, there are certain, you know, health outcomes, which are more prevalent among black people for, you know, genetic reasons, among other things. We, you know, that's, is that, is that something, you know, that's, that's in the real world, right? There's nothing bias happening. If we, if that, if that pops out and says, we need to be targeting more resources for, you know, for black people, because of these health issues. That's, that's because of a real world bias. Now, is that, is that bad? Should we be doing that? Like, or should we be attempting to ignore features that we don't think we should be taking account of? Potentially no

Speaker C: 00:09:33

less emotive though, is it? Just because it's a feature of the universe. I mean, that's the

Speaker B: 00:09:38

question. I think one issue here is, is going back to Peter's distinctions between the different kinds of biases to be clear about the question we're asking, right? And I think sometimes the, the sort of problem of bias, as you pointed out, is that we're, we think we're asking the question of, you know, who should get this healthcare treatment or who is likely to re-offend and who should get parole. But actually what we're asking is what did, what decisions did parole officers, parole officers make in the, in the past. And so being really clear about that is important. I also think, Nick, though you were getting at, perhaps like one thing I've been thinking a bit about is this, this bigger question that maybe goes beyond algorithmic bias, which is something about like what kinds of predictions we want to be able to make about people in particularly, like what kinds of ways we want to be able to, is it useful to be able to discriminate between people? So if we can use machine learning to better predict which people, like which subsets of society are more likely to get a disease or to go bankrupt, for example, like those predictions might be entirely accurate, but it might still be harmful to know these things because it's sort of, so I have a colleague who talks about this as kind of undermining this value of solidarity, like, and solidarity depends, this idea of solidarity in society, sort of treating everyone the same kind of depends on this veil of ignorance about who is going to get a disease. We have like, sort of, you know, social support systems that depend on some amount of ignorance. And that issue is even, is bigger than bias, right? We might be accurate, but.

Speaker A: 00:11:15

Yeah, I think that and that, that insurance is a great example, very controversial and a live issue. And you know, you can imagine getting to a sort of situation, I think I've heard people describe it as a kind of n equals one situation where you get so good at predicting someone's risk of something that you're more or less sorting them into one or zero, you know, you can sort of be almost perfect in your prediction about what's going to happen to someone and price insurance accurately. Well, that's controversial. I mean, no one, I don't, I think it's legal to charge women less for car insurance, because that's definitely a thing, right? I don't know if that's been successfully challenged.

Speaker B: 00:11:51

No, it's been, it's not, I think it's illegal now. I think it used to be legal, but it's been made.

Speaker A: 00:11:56

Okay. I mean, obviously, as an economist, I'm all in favour of pricing things accurately, because as soon as you obscure that information, you know, you're going to get distortions in the market, and you will no longer have an efficient level of insurance. But at the same time, as a human, it's incredibly annoying to be, you know, to be given a high price for something which is totally beyond your control. You know, where someone else who happens to live in, you know, Surbiton gets a nice, gets nice cheap insurance, you know, I can't help it, I can't help where I live, or, you know, what the fact that I'm male, it seems kind of unfair. So I think that's what we want to get Is it right to use the feature to try and use those features? Because we can expect us to be really good at learning to be discriminatory? Yeah, it's gonna learn really, really accurate ways of discriminating.

Speaker B: 00:12:49

That's quite useful, right? Like discriminating in the broad sense.

Speaker A: 00:12:51

But yeah, so, so, yeah. So I mean, what are some of the policy of you? I don't know if you've looked at any of the kind of what sorts of policy proposals people have come up with? Because I feel like, you know, we're completely on the back foot here. I mean, there are, because I've had a quick look at the legal side of this. Yeah. Legally, you know, there are these protected characteristics, which we might call features in our sufficient service context. And they are nine of them, which are age, disability, transgenderism, marriage, pregnancy, race, religional belief, sexual orientation, etc. You're not allowed to make different decisions on those bases, bases about people with a bunch of exceptions, like monks, you can't be a Muslim monk, not allowed. I suspect you can't be a Jewish imam. But anyway, you're allowed to discriminate on that basis. And actors and models in film, you've got a Chinese character, you you're allowed to say we want a Chinese actor to play them. And then you are allowed to target a sort of ethnic minorities if they're underrepresented for training schemes and that kind of thing. Things like cultural sensitivity. So you might want to employ, you know, a Muslim woman to do searches at airports and so on. Yeah, so there's a bunch of exceptions there. But I mean, so we have this sort of legal, legal setup, which says you can't, by and large, you can't use those features, not allowed. So you might think, well, we just won't put them into our model. We, I mean, one of the

Speaker B: 00:14:25

issues there is that often the model will still essentially learn these features through their correlations with with other things. One thing that's quite interesting there on the legal side of protected characteristics. So I was just actually listening to another podcast with Hannah Wallach, who's a researcher at Microsoft Research, who's been working in machine learning and fairness and machine learning for years. And she was talking about, I think it's this sort of conflict between what the law says you can do with protected characteristics and what technical approaches to ensuring fairness in machine learning actually need to do. So one sort of thing that she was talking about is that it's not clear whether protecting these characteristics means not using them in your analysis or like explicitly using them in order to make sure that you are treating people fairly across these these groups. And so it's like, should we be more aware of these

Speaker A: 00:15:14

characteristics or less? Just a quick question. So you sit out just outside of policy making, you consult, right? Yeah. I'm just wondering just how receptive or what's the kind of general knowledge within policymaking of the sorts of material that you're finding and how receptive

Speaker B: 00:15:34

are they to it? Yeah, good question. I mean, I think as is the case for like all policy areas, there is this gap between policy and academic research with problems on both ends, right? Like the academic research isn't always that relevant for policy and policymakers aren't that engaged. I think because AI and machine learning is so high profile and topical at the moment, there is more engagement on the policy level with what's going on in academic research. And so the UK government has just set up or is just in the process of setting up this new Centre for Data Ethics and Innovation, which will essentially, as I understand it, sort of try and advise, they're not going to be a regulatory body, but they're going to advise on regulation and policy related to AI and data. And they are like at the moment, commissioning academics to review research on algorithmic bias and communicate it to them and then start thinking about what that means for policy. So maybe sometimes I feel like there's too much hype around AI at the moment and it can be really harmful. But one of the good things about that is I think that governments are sort of having to take it really seriously and having to really engage with the academic communities.

Speaker C: 00:16:43

So on a practical working level, do you have any ideas of how you might sort of encourage your data scientists to be fairer, to be aware of their biases? So if you were advising this unit, what would you suggest that they do?

Speaker B: 00:17:03

So if I was advising the unit on what policy they need to do or how they affect government

Speaker C: 00:17:11

policymaking, so what advice would you say you need to give these departments this advice to

Speaker B: 00:17:16

help their data scientists? Yeah, I mean, as I say again, actually, so I know that this unit are thinking about algorithmic bias. And then the other thing they're thinking about, which relates to the sort of bigger thing we're talking about is basically sort of AI driven targeting and manipulation. And the targeting point is this, like, we can use machine learning to better understand differences in populations and therefore, like, give them different messages or give them different interventions and think about that. And that's actually the thing that I've been thinking more about than the bias. And I think, I mean, the sort of issue there is, we don't quite understand enough yet about what's being done to really understand what the issues are. But I think if we do need serious thinking about, so the GDPR exists and is a thing that I think is doing some work on sort of regulating the way that data and algorithms can be used to discriminate against people. And some of that relates to algorithmic bias, but I think there's a need to look more, and I'm not a legal expert, so I don't know this, but look more at what does the GDPR do in terms of constraining how information about people can be used to then deliver them sort of like targeted, tailored messages and services and what might be the harms of that. And is there a role, I think there might be a role for regulation to play beyond what the GDPR already does in kind of saying, like, you can't use certain kinds of information about people to determine which, you know, like, which policy intervention they receive or which...

Speaker A: 00:18:57

I'm going to come right out and say I think it's totally unenforceable. I think it's a much harder problem than I think these policy people are expecting.

Speaker C: 00:19:04

I mean, data science is a relatively new discipline, and data scientists are sort of inherently at the moment quite mercenary. I mean, they're employed by ad agencies to increase the number of clicks and by Amazon to increase the number of sales, and it's their job to just do that whatever. And although they might be conscientious in their jobs, they're under that restriction. If they don't do their job as well as they could, they'll be replaced. So I think that's always going to be a problem.

Speaker A: 00:19:34

Yeah, but I mean, I think we keep assuming that there's something bad about this. That's what I'm questioning. I think it's not even necessarily like good or bad. It is a completely necessary feature of what data analysis involves, is identifying features that meaningfully distinguish between things. The only way you can be genuinely unbiased about these things is by ignoring them, is by not factoring them into anything you're doing. So it's not possible to eliminate bias. Even if you take out gender, if it actually is affecting the world in some way, the machine will just learn proxies for gender. It'll learn proxies for race. It will learn about the things that predict what is going to happen to someone. There's nothing we can do about that. I cannot see how you, you know, then you might start saying, well, what we need to do is cut out anything which is correlated with gender or race or anything. Well, you can't. Is he accurate in saying this, do you think?

Speaker B: 00:20:43

So I have two thoughts on this. One is that, I mean, I think it's often being said now that a lot of the sort of ethical issues related to AI that people are talking about are not actually, necessarily novel and unique to AI, right? Like racial bias is not a unique problem for machine learning algorithms. It's a problem that's existed, you know, in society for yeah, for a really long time. And in a way, what our use of machine learning is doing is just putting this really clear spotlight on other problems. And so it might be the case that like, yes, it's not currently really possible for us to use machine learning in certain domains without coming up against these really big issues of bias. But what we can do is try and think about this as like, okay, this is alerting us and giving us a better understanding of bias historically, and then really creating like a stronger incentive to do something about that, like change the data or something. My second thought is that sort of going to this bigger picture question of like, what it is that we're using AI and machine learning for, it might be true that like, we can't use machine learning in advertising, say, without risking treating people unfairly or undermining their autonomy in some way. I worry a bit that, you know, we have these advances in machine learning and AI and getting lots of people excited and saying, you know, these could be used for a huge amount of good, but also have potential for harm. And, and one of the questions that's not being talked about that much is just like, well, what do we we have some, to some extent, limited resources, right, and limited ability to apply this stuff, like, what kinds of problems do we want to apply it to? And at the moment, a lot of the sort of jobs in AI and advances and stuff are going on in, say, advertising, right, because that's where the money is. But yeah, I don't know, maybe it's a

Speaker A: 00:22:40

little bit too blue skies. If it means I see fewer adverts for Stanna Stairlifts and things, that'd

Speaker C: 00:22:46

be great. Those walk-in baths. Well, you'll get a few more in a couple of decades time. I know,

Speaker A: 00:22:50

that's the worry, isn't it? No, because Facebook sent me a message about incontinence pants on my 40th birthday. It was the first time I'd ever seen an advert for it. I thought, oh God, is this the future? Yeah. Although, wasn't there a woman who got sent pregnancy vouchers? Before she knew she was pregnant? Yeah, there's stories like that, aren't there?

Speaker B: 00:23:15

Certainly before her dad knew she was pregnant. We're not at the stage of needing to wrap up yet,

Speaker A: 00:23:20

but we're sort of starting to glance in that direction. I've got one question I want to ask Jess, but I want to leave that for a while. We've got Jess. What else would you like to ask or to delve further into this? My question is, what do people think is bad about using all the information you can to make as good a decision as possible? Okay, if the data is biased, fine, but if it isn't, if we've got accurate measurements about the world and we are able to make accurate decisions and allocate resources, and if it turns out that that means more men than women are incarcerated and more healthcare resources are given to certain ethnic minorities, well, so what? That's good. Change my mind. I think one part of the worry is that this is just

Speaker B: 00:24:13

concentrating power, right? So yes, and potentially reinforcing existing power asymmetries and biases and things. So like, yes, machine learning enables us to potentially make better decisions and make better use of resources. But how machine learning is being used to do that is being decided by a very small group of people who have very specific sort of biases and ways of seeing the world. And so I think the bigger issue then, like this algorithm is biased, is this sort of power issue of like, these technologies are very, yeah, are very powerful data. I mean, data is incredibly powerful. We're seeing just like how having access to people's personal data is really concentrating power in the hands of these few tech companies. So I think that's the bigger

Speaker A: 00:25:08

issue people are worried about. It's more what you're saying is more, it's not so much that, okay, tactically, yeah, there's nothing wrong, better targeted adverts, maybe you're more likely to buy the right incontinence pants. But actually, the problem is that everyone is doing advertising and perhaps we, the fact, yes, we could use these things to target healthcare resources. But our fact is we're not and we should be. So what things, what are the unsolved problems that you think we need to be throwing compute at? Yeah, no, don't ask me that question. No, go on, I'm afraid it's out there now. If you had to pick one. I mean, there are, yeah, there are. So I'm going to

Speaker B: 00:25:43

go on. So I was at, I was at an event the other week where someone asked the question, I think we need to be asking, or like, what would AI systems look like? And how would they look different from the way they look today? If we were really trying to use them to solve the world's problems? And I suppose the really worrying thing is like, I don't automatically, I don't know how to answer that question. And like no one in this room, there are a bunch of experts really knew how to answer that question. I mean, I think, you know, there are there are there are applications that are exciting. I think, I think healthcare is like an obvious domain where we do have a lot of data and actually better understanding patterns in the data and relationships could clearly be very valuable. It's also a domain where we really have to deal with like privacy issues and where discriminating between individuals might be harmful. But I do think it's an area that has a huge amount of promise. Yeah. I mean, I don't think that's most important. It's a bit of a painfully

Speaker C: 00:26:47

hypothetical question, really. It's like, how? The fact that a lot of AI... All we know is the answer

Speaker A: 00:26:53

isn't advertising. Yeah. I mean, I think the way to... Marketing probably isn't the best way to

Speaker B: 00:26:58

make the world a better place. And it's a very high level way to think about what machine learning helps us with, is it does just help us to better understand really complex patterns in the world that like our human brains can't intuitively understand. And there are tons of things there that could be really useful, right? Like better understanding like climate dynamics and figuring out what might help us to solve climate change. Yeah. Yeah. I mean, I don't know, the economist

Speaker A: 00:27:28

in me wants to say, well, the problem of scarce goods and things being expensive and logistics being costly affects everyone. If you have to pay more for food because it's inefficiently brought to Britain rather than through a really efficient network, well, it's good. It's good that companies are able to cut costs. It's good that advertising is able to pay for the free internet that we have. It's good that I get adverts for incontinence pants because people like me are more likely to be incontinent. Maybe that's a good thing. I'm not. That's because you've got the pads on, that's why.

Speaker B: 00:28:09

But I suppose there's being good and it's like, is this the best use of unlimited resources, right?

Speaker A: 00:28:15

When I asked you what the best use was, you didn't know, right?

Speaker B: 00:28:18

No, no, I don't. But I can think of things that are better, right? Like I've been talking to a lot of people about applying AI to agriculture and I don't really understand agriculture, but it seems like there's a bunch of potential to improve food security, for example. And there are risks, again, associated with that. But that seems better than making some company more efficient or...

Speaker A: 00:28:40

But presumably, even if we do use it for agriculture, it will make some company more efficient. I mean, it's going to make a farming company better at doing that. And I guess, I don't know,

Speaker C: 00:28:50

it might make them more profitable, but not necessarily better for the planet. Yeah. And we need to sort of draw to a close. Peter, was there something you wanted to add? Well, yeah. I mean, just to this, there is some silver linings. There's certain things I've got to mention as a rule in our podcast. Oh, God. Blockchain? So, yeah. Well...

Speaker A: 00:29:10

Oh, no. He said it. The blockchain cat is out of the bag.

Speaker C: 00:29:13

Well, no, it's related to a principle that can be enabled by blockchain technology. So a lot of research being done by Microsoft and Facebook and the big tech giants into new ways that people can manage their own data online to authenticate themselves. And there's a principle called self-sovereign identity, which is particularly exciting, that the sovereignty is putting your data back into your control. So you're self-sovereign in the sense that you have control over your own identity. And that opens up lots of possibilities in terms of privacy and preventing inappropriate data access, but also means that people can be directly remunerated for use of their data. So rather than, as we do now, kind of recklessly hand over control of our data just to get that new app on our phones, means that the people who want to subsequently use your data is much more transparent to you as an individual. So you have choice. It's much more democratic about which models you elect to be part of and which ones you don't. So it's just a technical...

Speaker A: 00:30:30

Is that a statement or a question?

Speaker C: 00:30:31

No, it's a statement. There's things such as this which are potentially part of a solution.

Speaker A: 00:30:40

Okay. I mean, just to round things off, Jess, can you tell me why it is you do what you do? And I guess part of that is what is it that excites you about your field and what could happen in the future?

Speaker B: 00:30:52

Yeah. So it's interesting because I sort of came to thinking about issues around AI from two different directions. One was that I had this background in behavioral science and I was really interested in the way that humans make decisions and the sort of flaws in that and the way that those kind of flaws in decision making contribute to problems in the world. And I was really interested in how improving human decision making might actually help us better solve problems in the world, but got quite disillusioned with all these sort of ideas of de-biasing humans. And so as a result, got much more interested in like, well, maybe these AI systems, which have very different strengths to humans, might help us to solve some of the complex big problems that we can't solve ourselves. So that's the sort of more positive side of like being excited about these systems that have very different strengths to humans and might be able to complement us in various ways. And then the more negative side was being, I suppose, worried, like worried about, worried on multiple levels about what's very advanced AI systems that are much smarter than humans might mean for the world, but also worried about what's the sort of concentration of power that these systems bring in happens to society and, and all of those things. So there's kind of these dual motivations, like I'm motivated by the potential for this technology, but also by, and a lot of the time at the moment by wanting to make sure that we kind of, yeah, prevent the harms.

Speaker A: 00:32:22

Nice answer. Just rounding things off, might be a stupid question. Yeah, no, we always have to end on a stupid question. So it's got nothing to do with AI. Perfect, brilliant. Nothing to do with what we've been talking about at all. Relevant as always, it's great. But I was just... What's your favourite cartoon? We can get your ideas. I was just, I was just imagining if you were a party, right? And let's say it's a, I'm sure you go to loads of parties with loads of like data analysts and ethicists and so on. But let's say this one isn't one of those. It's a party with normal people. And someone goes, oh, what do you do? And this is going to be for all of us. Oh God, yeah. And so what do you say to that person in one sentence? So this is what, this is my job, this is what I do. That's question part A. Okay. And question part B is what do you think you'd be, what would you do if you weren't, either what would you want to do if it's not what you already do? Or what else? What's your other career that you're not doing? Okay. Good questions or bad questions? It might tie in somehow. So anyway, you don't have to go first, Jess, actually. Actually, no, I think you do.

Speaker B: 00:33:31

Okay.

Speaker A: 00:33:32

Yeah, go for it.

Speaker B: 00:33:35

So interestingly, I think my, my answer to what you do depends on the person I'm talking to and sort of what I feel like on the day. Because I don't feel like it's very easy to explain exactly what I do. But my default is I do research into a range of ethical and policy issues related to artificial intelligence. Sometimes I'll just say I'm an academic, or I'm a researcher and then see if people prod me anymore.

Speaker A: 00:34:01

I think that's reasonable. And actually, before we go around to part B, Peter, what would you say to someone at a party?

Speaker C: 00:34:06

Well, if I'm being obtuse, I might say that I'm a conceptual engineer.

Speaker A: 00:34:13

Strangely, he doesn't get invited to any parties. I know what I would do at that party.

Speaker C: 00:34:16

What I really mean is that we do sort of research and development stuff. So we help formulate ideas and approaches to achieving certain aims.

Speaker A: 00:34:29

That's nice. Yours must be the same, or how would you describe it? No, I normally go with, I'm the director of an analysis and decision support company, and we help people make better decisions with the data that they have. I like that, apart from the decision support bit. I mean, that's, don't know. Well, we support decisions. I mean, what's that? It's not decision support, you tell me. No, but I think there's a nicer way of saying support. There's a more, yeah, I don't know. Mine's really simple. I make films. Simple as that. And record the odd podcast, you know. Okay, so bringing it back round. What would you be doing if you weren't doing this, Jess?

Speaker B: 00:35:04

I want to have an answer that's like really out there and fun and crazy, but I don't. So I mean, I think a lot of what I like about research is actually writing. I think part of me is more of a writer at heart. I think part of the reason I'm not a writer is it's just a much harder career path to go down, and I like the research part, but there's a part of me that I feel like I could have been a writer or a journalist or something.

Speaker A: 00:35:28

Okay.

Speaker B: 00:35:28

Maybe.

Speaker A: 00:35:29

It could be your side hustle.

Speaker B: 00:35:31

Yeah, maybe one day, maybe one day. I have dreams of, yeah, just quitting things and sitting and writing a book or something one day.

Speaker A: 00:35:38

In one of those swinging wicker chairs. Lots of cushions. Yeah, I can totally see that. Peter or Nick?

Speaker C: 00:35:47

Well, I've got lots of grand, high-minded fantasies about, you know, having had a successful career in the army and potentially being a British astronaut and all that sort of stuff. But I'd actually probably most likely just be a boring, broken, grey civil servant still.

Speaker A: 00:36:01

Peter, that's not... I think you've missed the spirit of the question a bit there. You can take it as you like. No, no, that sort of answered both sides of it, actually. Nick? I really love what I do. So I don't think, it's very hard to imagine a different thing that I would be doing that I would enjoy more. But I think it would be the same, but perhaps less concerned about actually earning money. I'd be doing similar things. Again, yeah, research, probably. I'd be doing a lot more. I'd probably spend more time designing and running games and that those kinds of sort of soft analytical things, which I enjoy, you know, writing, probably like Jess, we're doing a bit more forecasting, a bit more writing about the future, you know, probably a bit more speaking at things. It's a bit, yeah. I mean, the thing about the job, which is, you know, makes it a job is having to sit down and research stuff you're not particularly interested in and come up with insights on behalf of a customer. I'd probably, that's the bit I'd shave off, I think. Yeah, yeah. Were you going to add to something

Speaker C: 00:37:02

else? Yeah, second bite of the cherry. I would quite like to do stuff with my hands. I'd quite like to build things. I get great pleasure in actually achieving something physical. So if a second career after this would be to set up a sort of prototyping and creative space full of raspberry pies and oscilloscopes. The cog space, be called the cog space. And yeah, and just as a space for

Speaker A: 00:37:31

people to come and use and things. So my answer is actually funny enough, the cross between both of you, because what I do is actually what I really like, but I'd like to do it without sort of commercial necessities. And I'd love to be off photographing stuff, filming stuff, and not just having to do what I do. But so, so I'm kind of almost sort of there. But also I do have a second career, well it's about a fifth career actually, at some point, which is I want to build boats and I want to build wooden boats and do something with my hands. And I have this sort of idyllic thing. I live in some community somewhere, small, down by the sea. Wooden ships and iron men. Yeah, exactly. And I'm, I go into the power, all right, Fraser, the boat builder, coming out of the boat. And so anyway, that's my sort of- Yeah, the thing is about you, Fraser, is you'll probably bloody go and do it. I mean, that's, that's the thing. That's your problem, you see, is you still haven't quite got the hang of the difference between your real self and your idealized self. I mean, most people come to terms with that in their early teenage years, but you're, you're still getting

Speaker C: 00:38:32

there. Let's go, let's go and get a, let's get a good old shipwright yard and I'll set up my workshop. We could make this happen. Jess, you in? You can make the sails. The other thing I was

Speaker B: 00:38:42

nearly going to say is I feel like there was, there was a time when I could have been an architect. Like I liked maths and I like this sort of like precise drawing, so maybe I can bring some of my, my architecture skills to the, can design the boats and you can build them. Naval architect?

Speaker A: 00:38:56

There's something beautiful going to happen here. I can feel it. Okay. We're going to have to draw to a halt there. Suffice to say, thank you as always for listening to the Cognitive Engineering Podcast. You've been here with myself, Fraser McGruer, with Nick Hare and Peter Coghill of Aleph Insights, but most of all, thank you very much indeed to our special guest, to Jess Whittlestone for being with us today and for sharing all your insights. So thank you very much indeed.

Speaker B: 00:39:17

Thank you for having me.

Speaker A: 00:39:17

Thank you. Goodbye.